Table 4 Performance of the CDACNN and DenseNet in the internal test dataset
One-vs.-Rest Classification | CDACNN | DenseNet | ||||
|---|---|---|---|---|---|---|
Sensitivity (95% CI) | Specificity (95% CI) | Accuracy (95% CI) | Sensitivity (95% CI) | Specificity (95% CI) | Accuracy (95% CI) | |
Keratitis vs. other + normal | ||||||
 Conj_Cor | 98.6% (0.975–0.996) | 99.2% (0.984–1.000) | 98.9% (0.982–0.995) | 97.1% (0.956–0.986) | 99.8% (0.994–1.002) | 98.5% (0.977–0.992) |
 Cornea | 98.3% (0.972–0.995) | 98.2% (0.971–0994) | 98.3% (0.975–0.991) | 98.1% (0.969–0.993) | 96.2% (0.946–0.979) | 97.2% (0.961–0.982) |
 Original | 97.7% (0.964–0.991) | 97.2% (0.958–0.987) | 97.5% (0.965–0.984) | 97.9% (0.967–0.992) | 96.4% (0.948–0.980) | 97.2% (0.961–0982) |
Other vs. keratitis + normal | ||||||
 Conj_Cor | 96.9% (0.940–0.999) | 99.2% (0.986–0.998) | 98.9% (0.982–0.995) | 99.2% (0.977–1.007) | 98.1% (0.972–0.990) | 98.3% (0.975–0.991) |
 Cornea | 91.5% (0.868–0.963) | 99.1% (0.984–0.997) | 98.1% (0.972–0.989) | 88.5% (0.830–0.940) | 98.7% (0.980–0.995) | 97.4% (0.964–0.984) |
 Original | 85.4% (0.793–0.915) | 98.6% (0.978–0.994) | 96.9% (0.958–0.979) | 84.6% (0.784–0.908) | 98.5% (0.977–0.993) | 96.7% (0.955–0978) |
Normal vs. keratitis + other | ||||||
 Conj_Cor | 100% (1.000–1.000) | 100% (1.000–1.000) | 100% (1.000–1.000) | 99.5% (0.987–1.002) | 100% (1.000–1.000) | 99.8% (0.995–1.001) |
 Cornea | 100% (1.000–1.000) | 99.7% (0.992–1.001) | 99.8% (0.995–1.001) | 98.1% (0.967–0.995) | 99.8% (0.995–1.002) | 99.2% (0.986–0.997) |
 Original | 99.5% (0.987–1.002) | 99.0% (0.982–0.998) | 99.2% (0.986–0.997) | 98.7% (0.975–0.998) | 99.3% (0.987–1.000) | 99.1% (0.985–0.997) |